Add FIRMCore evaluation task implementation

[pardiszahraei]

Adds a task registration for the FIRMCore fairness/bias benchmark to olmo-eval.
FIRMCore is a 50,012-item core subset of the larger FIRM benchmark, selected via
Fluid-Benchmarking-style IRT-based item selection, covering ground-truth capability
items and injected-pressure bias-lean items across ~29 domains and 8 demographic
pairing categories. This PR adds firm_core.py, registering firmcore as a
runnable Task sourced directly from the allenai/firmcore-benchmark HF dataset
via DataSource (currently private, scheduled to go public within about a month).

Includes:

• 4 metrics, named for what they measure rather than abbreviated
  (decision_accuracy_gap, minority_selection_rate_tied_cases,
  correct_decision_flip_asymmetry, injection_induced_selection_shift).
• Answer extraction robust to per-domain role-noun labels (e.g. "Patient A", not
  just "Candidate A"), an explicit "Answer: X" line, and candidate display-name
  matches (full or partial, e.g. a bare first name).

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

No pytest files were added to this repo for this task specifically. Instead,
verified manually:

• uv run olmo-eval task inspect firmcore -- confirms the task loads and
  instances format correctly
• uv run olmo-eval run -m mock -t firmcore -o limit=200 -- full
  data-loading -> formatting -> scoring -> metric-aggregation pipeline runs
  end-to-end without errors
• Standalone script testing the answer-extraction function and all 4 metric
  classes against hand-crafted responses with known, hand-computed expected
  values (all matched)

Happy to add a proper pytest file under tests/ if that's the convention here.

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my code
• I have added/updated documentation as needed
• My changes generate no new warnings
• Any dependent changes have been merged and published (N/A, no dependent changes)